Your achievements

Level 1

0% to

Level 2

Tip /
Sign in

Sign in to Community

to gain points, level up, and earn exciting badges like the new
Bedrock Mission!

Learn more

View all

Sign in to view all badges

Gaining a Deeper Understanding of Churn Using Data Science Workspace

Avatar

Administrator

06-10-2021

Authors: Marina Mahtab, Sunish Verma, and Douglas Paton

Banner image.jpeg

In this post, we explore how we helped a customer better understand churn within their company using the DataScience Workspace feature of Adobe Experience Platform. We break down the process and explore the attributes of customer churn.

In highly competitive markets, like gaming, hotel, and casino chains operate in, being able to truly understand your customers is crucial to remaining competitive. To do this, you have to analyze the behavior of your customers and understand what are the attributes of people who may churn.

With that in mind, we wanted to use Data Science Workspace to leverage machine learning to predict with a high accuracy which customers of the gaming, hotel, and casino major were highly unlikely to make a booking over a six month period. Then, using those predictions, we would send personalized offers to customers through either Adobe Target or emails, as well as social media. We used Adobe Experience Platform for this because of its appetite for data. Using Data Science Workspace and Adobe Experience Platform together allowed us to be able to activate and report within one tool.

At a high-level the below architecture involving data science workspace helped us deliver the insights on churn.

Figure 1: A high-level look at how the data gets analyzedFigure 1: A high-level look at how the data gets analyzed

We looked at customers who had booked with the chain any time during an 18-month window, we discovered that 95% of customers re-book within six months. This meant that, with a six-month analysis window, we could understand one complete purchase cycle for customers.

Our data included half a million customers and with a churn rate of 53%.

We explored churn attributes for each customer such as the number of orders, days since order compared to the customer’s average buying cycle, customer persona as evident from browsing behavior such as sensitivity to price, sort by popularity, seeking deals as well as other session-related activities that flow into the platform in XDM format. We then leveraged the data science workspace to clean out the data and select the meaningful drivers of churn for building the predictive model. Recipe feature of the data science workspace was used to create, experiment, and tune ML models right where data is collected and activated, shortening data science time to insights.

The ultimate goal for our analysis is to understand the attributes that impact churn and determine a propensity to churn over the next few months.

Figure 2: Model lifecycle — From features explored to action in segmentsFigure 2: Model lifecycle — From features explored to action in segments

Insights for sharper targeting

As we explored the data, we found that there were certain factors that were common among those who churned.

  • The distinction between churners and non-churners becomes sharper if we consider only the recent months. So a customer who made frequent bookings in a one month period was more likely to churn out than when we expanded our view out to several months. Consistent engagement of the customer over longer durations can help prevent churn.
  • The properties that customers visited impacted whether or not they churned. For example, customers who visited certain properties had a higher chance of churning. Similarly, there were also some properties after a stay at which customers were more likely to return to make a booking.
  • Interestingly enough, gold tier members were more likely to churn, as were those who had received a high number of comps from the customer reward program. There was among a tendency those who churned to search for properties they wanted to stay in based on price, sorted low to high.
  • Customers who arrived via bookmarks (typed_bookmark as the referrer) had an increased likelihood of churning, as were customers who had “booking confirmation” as their original entry site.
  • We also found that customers who left at the “booking confirmation” page often never returned to the site.

Model

To combine all the insights into a single usable metric we built a machine learning model to predict churn as soon as the first signals of churn are exhibited by the customer for sharper targeting and retention.

For this we leverage the XDM data model and Jupyter Notebooks in Data Science Workspace in Adobe Experience Platform. Adobe Experience Platform also has pre-built notebook templates and a number of useful plug-ins including a one-click notebook-to-recipe packaging extension, a service for scaling the multi-tenancy of Jupyter, and parameterized notebooks.

Designing the model

In simple terms, machine learning models help in predicting an outcome based on historical patterns. For example, by reviewing the online advertisements that got the maximum clicks in the past, you can predict what is that particular feature that makes a user click instantly.

To discover these patterns for churn, we trained the model on customer behavior for one purchase cycle, as discussed above is nearly six months. Using this as a benchmark we designed our model to predict a customer would return to make a booking in the next six months. We also studied the data across time periods to make sure there was no major seasonality in orders; seasonality in data can influence the model performance during peaks or troughs.

Training the model

There are a plethora of machine learning algorithms to choose with each its own applicability, we tried several algorithms like logistic regression, random forest, support vector machines, and neural networks, we chose artificial neural networks for operationalization as it worked best based on three different criteria:

  1. Recall of the model: The number of churners the model is able to identify and cover
  2. Precision of the model: The number of correctly identified churners as a percentage of the total predicted as churn
  3. Stability of the above metrics over different time periods

Neural networks are a set of algorithms, modeled loosely after the human brain, that is designed to recognize patterns. Here, we trained the neural network to catch the first signals of churn.

Some challenges and next steps for us

In order to make sure that we achieved a single view of customers that combined data and churn predictions for customers, we had to undertake a very detailed analysis during the feature engineering stage. Featuring engineering involves creating meaningful and compelling features (a customer characteristic that is being observed in data as a precursor to churn) and requires a critical eye to select the best features for predicting churn.

When we started, we had over 150 features from the Analytics (Clickstream) data aggregated as sales and commerce, web behavior, past response to marketing, customer demographics among others. Through careful analysis and consideration, we were able to reduce this down to 20.

Our model building efforts here can be deployed as a service to help customers understand and control churn within their company on a continued basis.

What’s great about all of this is that we can use this to analyze behavior. This allows us to create Adobe Experience Platform Real-Time Customer Profiles that we can use to determine whether or not a customer will churn. And put us in a position to get highly-targeted, personalized offers out to them as soon as they start showing signs of churning. We can use this data for reporting and giving insights to the organization leveraging the likes of customer journey analytics and Query Services.

We’ll talk more about those Real-Time Customer Profiles in an upcoming blog post. You can learn more about model deployment within Data Science Workspace here.

Follow the Adobe Experience Platform Community Blog for more developer stories and resources, and check out Adobe Developers on Twitter for the latest news and developer products. Sign up here for future Adobe Experience Platform Meetups.

References

  1. Adobe Experience Platform — https://www.adobe.com/experience-platform.html
  2. Data Science Workspace — https://www.adobe.com/experience-platform/data-science-workspace.html
  3. Experience Data Model — https://www.adobe.io/open/standards/xdm.html
  4. Adobe Experience Platform Real-Time Customer Profiles — https://www.adobe.com/ca/experience-platform/real-time-customer-profile.html
  5. Jupyter Notebook — https://jupyter.org/

Originally published: Jul 9, 2020

1 Comment